ITL Ideogram 4 Prompt Builder

Ideogram 4 is a strange model: it's trained almost entirely on structured JSON captions, not prose, so a plain prompt like "a red fox in a forest" comes out noticeably worse than the same idea written as fields, hex palettes, and bounding boxes. The catch is that the schema is miserable to type by hand. This node is the fix - a full visual canvas editor where you draw boxes, describe each one, and it assembles the caption JSON for you.

The name might make you expect an API key and a server call. It's nothing of the sort: this is a local ComfyUI node that just formats text. The heavy lifting happens in the model you wire the output into - Ideogram 4's own sampler - and this node only builds the JSON that model was trained on.

How it works

Ideogram 4's caption format centers on a compositional_deconstruction with elements - each one has a bbox, a desc, an optional verbatim text, and a color_palette of hex colors. Boxes live on a 0–1000 grid as [ymin, xmin, ymax, xmax], they can overlap and nest, and that's the whole trick: the model gives you genuine regional control because boxes were baked into its training, not bolted on afterwards (the regional-prompting story is long; Ideogram 4 is the first time it's native).

You draw on the canvas, the editor serializes your regions into the hidden elements_data field, and at run time the node merges that with your text fields and outputs one compact JSON string. If this smells like Kijai's KJNodes Ideogram builder, that's because the author lifted the canvas editor from it (with attribution, GPL-3.0) and layered on parenting, named groups, and a scene-tree overview - but it coexists fine with KJNodes, so you don't have to uninstall anything.

The inputs that matter

Most of the node's inputs are managed by the UI - elements_data and style_palette_data are just serialized editor state you shouldn't touch. The ones you actually set:

• width / height - the canvas aspect (and the pixel grid boxes are measured in). Ideogram 4 wants multiples of 16, and the node enforces that with its step of 16. Default 1024×1024.
• background - the required one-line scene description that sits behind all your boxes. It's marked required for a reason: leave it blank and you get a meh image.
• high_level_description - optional one-liner overview of the whole image; blank omits it.
• style (plus aesthetics, lighting, medium) - Ideogram's style descriptors. These are where the companion ITL Style Wizard plugs in.

Two that bite beginners: import_mode controls what a wired import_json does - when empty only seeds the editor (so you can still edit), always makes the wired JSON authoritative. And the image input is just a reference shown as the canvas background so you can trace boxes over a real picture - the "Grab BG" button does the same with your last generated output.

Outputs

• prompt (STRING) - the caption JSON. This goes into whatever node your Ideogram 4 workflow uses to encode the structured caption.
• preview (IMAGE) - a rendered overlay of your boxes, so you can eyeball the layout.
• bboxes (BOUNDING_BOX) - pixel-space boxes, handy if you're feeding a SAM3 or crop step downstream.
• width / height (INT) - the canvas dimensions, for Empty Latent or resolution plumbing.

Installing it

Grab it through ComfyUI Manager (search "ComfyUI-IntoTheLatent-Utils") or:

cd ComfyUI/custom_nodes
git clone https://github.com/Into-The-Latent/ComfyUI-IntoTheLatent-Utils

Then restart ComfyUI. No extra dependencies - it uses Pillow and the canvas is pure front-end JS, both already shipped. Note the repo is published for reference: issues are disabled and PRs are auto-closed, so don't expect support.

Gotchas

The one I keep hitting: the canvas editor lives in the node's own panel, and new users sometimes edit a region, close the panel, and wonder why nothing changed - the node's output only refreshes on a run, and output_format/coord_mode/bbox_order are all toolbar settings, not widgets. Leave coord_mode on normalized (the 0–1000 grid Ideogram expects) and bbox_order on yx unless you're exporting for Qwen-VL, which wants xy. And remember this pack renamed nodes from the old AI2Go* IDs - old workflows won't auto-pick these up, but keeping the old pack installed lets both coexist.